As we expected, the first products using ASICs co-packaged with optics were announced in 2021. Broadcom surprised the industry with a very aggressive timeline for Ethernet switches equipped with CPO in their January 2021 announcement. The company plans for demos of these switches in the next few months and actual products by the end of 2022. Broadcom’s plans have to be taken seriously. The company has a great track record for delivering. Yet, the industry remains skeptical.
The skeptics argue that the development of the co-packaged optics technology is just starting and real products are years away. COBO and OIF announced their working groups in the end of 2020 and have just started to define specifications. CPO collaboration launched by Facebook and Microsoft offered guidelines for suppliers in 2019 and the first draft of 3.2T CPO Product requirements in February 2021, but the organization has been suspiciously quiet since then.
It seems that the early excitement about CPO cooled off in 2021. Broadcom’s announcement may have raised concerns about the increasing dominance of this company in the switch market. Cloud companies spend billions of dollars a year on Broadcom’s switching ASICs and their priority is to increase competition in this market rather than having one supplier leading the innovation and controlling the market. Facebook emphasized the need to develop industry standards for CPO and have a new ecosystem with multiple suppliers of standard optical chiplets designed for switching ASICs, mimicking the very competitive supply chain for optical transceivers. Proprietary CPO solutions are not a preferred option for Facebook and other Cloud companies.
Amazon started to deploy switching ASICs made by Innovium (now part of Marvell) in 2020, to diversify its supply chain. Facebook announced the first switches using Cisco’s SiliconOne ASICs near the end of 2021. Google mentioned it is testing switches made by Intel. All these suppliers of ASICs are working on switches with CPO as well, but none of them has disclosed a product roadmap yet. Broadcom’s announcements added pressure on all the other suppliers to accelerate the development of CPO, and we may see more product announcements in 2022 and certainly in 2023.
While the skeptics may have dominated the CPO discussion in 2021, development efforts have probably accelerated since Broadcom’s announcement. It is true that CPO standards are years away and the first products will be proprietary, but if customers have a choice between several proprietary solutions, this may create enough competition among suppliers. Instead of creating an eco-system with numerous suppliers of optical chiplets, we may end up with an eco-system of numerous suppliers of switching ASICs with CPOs. Can new standards be developed on this level? Switch abstraction interface, developed by the open compute project, is clearly a step in this direction.
The dilemma between the standard and proprietary solutions is often summarized as: “if you want to go fast – do it alone, if you want to go far – do it together”. CPO is in the “do it alone” stage now, but the industry will need to start “doing it together to go far”. Both stages are necessary to bring a new technology to a broader market.
Public discussions are focused on co-packaging optics with switching ASICs, but CPO is a much higher priority for CPUs, GPUs and TPUs used in HPCs and AI Clusters. These systems are starved for bandwidth. A factor of 10x increase is needed now and another 10x will be needed for the next generation architectures based on disaggregated compute and storage.
As illustrated in the figure below, HPC and AI Clusters will be the largest market segment for CPO optics by 2026. We forecast CPO shipments in terms of 800G and 1.6T ports, which can be combined into 3.2T or 6.4T optical chiplets. This approach offers a clear view of transition from pluggable optical transceivers to CPO, which will be very gradual.
HPCs were the first to use massive quantities of EOMs back in 2010 and none of the EOMs made since then were standard solutions. This is certainly one of the reasons for limited use of EOMs and a lesson for CPO suppliers. The majority of EOM shipments in 2021 are going into military and aerospace systems, shown in the figure as “Other”. These are niche markets, which are too small for standards to develop. This is an example of “to go fast” strategy without a plan “to go far”.
The distinction between EOMs and CPO is blurry. CPO is really the next generation more compact EOMs, designed to be placed much closer to an ASIC, enabling co-packaging. While EOMs have largely failed to compete with DACs, AOCs and optical transceivers, CPO is expected to be more successful. Attaining another 10x increase in bandwidth, while limiting power consumption, is almost impossible without CPO technology.
More information on the report is available at: Market reports